BUG: bdate_range(end, periods, freq=nB) anchors stride at buffer start

[jbrockmendel]

Summary

Follow-up to GH-64648, addressing a regression flagged by @jorisvandenbossche in post-merge review.

When bdate_range / date_range is called with end + periods and freq="nB" (or "nC") where n >= 2, the new fast path in generate_daily_offset_range anchored the [::n] stride at the start of the internal buffer instead of at end. The result was offset by buffer_days % n:

>>> pd.bdate_range(end="2026-03-22", periods=3, freq="2B")
# main (before fix):  ['2026-03-13', '2026-03-17', '2026-03-19']
# expected / 3.0.x:   ['2026-03-16', '2026-03-18', '2026-03-20']

n=1 was unaffected. Only n>=2 with end+periods is broken.

The fix reverses the array before striding so the anchor is the last on-offset date <= end, then reverses back. Symmetric with the existing trim_from_end branch a few lines later.

This regression only exists on main — 3.0.x doesn't have GH-64648 and uses the old _generate_range path, which handles nB correctly. No 3.1.0 has been released yet, so no whatsnew entry.

Test plan

• New parametrized regression test covering end on Fri/Sat/Sun with freq="2B" and "3B"
• Full pandas/tests/indexes/datetimes/test_date_range.py (421 tests) passes
• Full pandas/tests/tseries/ (6941 tests) passes
• mypy and pre-commit clean

🤖 Generated with Claude Code

[jorisvandenbossche]

Thanks, looks good to me